Learn · Track
What the Model Context Protocol is, who talks to whom, what tools, resources and prompts are each for, how a tool call travels, which transport carries it, how remote servers ask for authorization, and what tool annotations can and cannot tell you. Written against the specification's own text, with the 2026-07-28 revision noted where it changes things.
MCP explained: hosts, clients and servers, the tools, resources and prompts a server offers, how a tool call travels, and a runnable agent to try.
The three MCP roles in plain terms: what the host app, each client connection and each server own, local vs remote servers, and where credentials live.
When an MCP server should offer a tool, a resource or a prompt, who decides each is used, and what output schemas, tool errors and elicitation add.
How MCP messages travel: stdio for local servers, Streamable HTTP for remote ones, sessions, Origin checks, and what the stateless 2026-07-28 spec changes.
How a remote MCP server asks for OAuth: the 401 challenge, resource metadata, PKCE, the resource parameter, and why tokens must never be passed through.
What readOnlyHint, destructiveHint, idempotentHint and openWorldHint mean, why the spec calls them untrusted, and how to use them in tool rules safely.
Prerequisites: none. Knowing what an HTTP request and JSON look like helps with the authorization guide.
After this track: MCP security risks covers what can go wrong with servers, and How AI agents work shows how agents use MCP tools in a loop. To build one, try the zero-key MCP agent recipe.
This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings